Papers by Yinfei Yang
Self-Supervised Learning for Pairwise Data Refinement (2020.aacl-main)
Copied to clipboard
| Challenge: | Pairwise data constructed from weakly supervised signals is widely used for training deep learning models. |
| Approach: | They propose two methods to refine pairwise data that are aimed to obtain subsets that are more useful as learning examples. |
| Outcome: | The proposed methods achieve most machine translation gains in the first iteration, but following iterations further improve its intrinsic evaluation. |
Improve Vision Language Model Chain-of-thought Reasoning (2025.acl-long)
Copied to clipboard
Ruohong Zhang, Bowen Zhang, Yanghao Li, Haotian Zhang, Zhiqing Sun, Zhe Gan, Yinfei Yang, Ruoming Pang, Yiming Yang
| Challenge: | Current training recipes often rely on datasets dominated by short annotations with limited rationales, hindering the models' ability to generalize to tasks requiring comprehensive reasoning. |
| Approach: | They propose a two-stage post-training strategy that augments short answers with CoT reasoning generated by GPT-4o, enhancing the VLM's CoT capabilities through fine-tuning. |
| Outcome: | The proposed strategy enhances the model's CoT capabilities through fine-tuning and reinforcement learning. |
Zero-shot Neural Passage Retrieval via Domain-targeted Synthetic Question Generation (2021.eacl-main)
Copied to clipboard
| Challenge: | Recent advances in neural retrieval have led to advancements on document, passage and knowledge-base benchmarks. |
| Approach: | They propose an approach to zero-shot learning for passage retrieval that uses synthetic question generation to close this gap. |
| Outcome: | The proposed approach can exceed term-based techniques on document retrieval benchmarks by using domain-targeted synthetic question generation. |
Neural Retrieval for Question Answering with Cross-Attention Supervised Data Augmentation (2021.acl-short)
Copied to clipboard
| Challenge: | Early fusion models with cross-attention have shown better-than-human performance on some question answer benchmarks, while it is a poor fit for retrieval since it prevents pre-computation of the answer representations. |
| Approach: | They propose a supervised data mining method to train an efficient late fusion retrieval model by using cross-attention models with cross-references. |
| Outcome: | The proposed model outperforms retrieval models trained with gold annotations on Precision at N (P@N) and Mean Reciprocal Rank (MRR). |
Learning a Multi-Domain Curriculum for Neural Machine Translation (2020.acl-main)
Copied to clipboard
| Challenge: | Existing data selection methods do not work well for multiple domains . multiple aspects need to be considered for training a multi-domain model . |
| Approach: | They propose a dynamic data selection method to multi-domain NMT that incorporates instance-level domain-relevance features and a curriculum to gradually focus on multi- domain relevant data batches. |
| Outcome: | The proposed model outperforms no-curriculum training on multiple domains and reaches or outperformed individual performance. |
Multilingual Universal Sentence Encoder for Semantic Retrieval (2020.acl-demos)
Copied to clipboard
Yinfei Yang, Daniel Cer, Amin Ahmad, Mandy Guo, Jax Law, Noah Constant, Gustavo Hernandez Abrego, Steve Yuan, Chris Tar, Yun-hsuan Sung, Brian Strope, Ray Kurzweil
| Challenge: | Using a multi-task trained dual-encoder, our models embed text from 16 languages into a shared semantic space. |
| Approach: | They propose retrieval focused multilingual sentence embedding models on TensorFlow Hub. |
| Outcome: | The models achieve state-of-the-art on monolingual and cross-lingual retrieval (SR) and retrieval question answering (ReQA) competitive performance is obtained on related tasks of translation pair bitext retrieval and retrieving question answering. |
Predicting Annotation Difficulty to Improve Task Routing and Model Performance for Biomedical Information Extraction (N19-1)
Copied to clipboard
| Challenge: | Modern NLP systems require high-quality annotations, but experts are expensive and lay annotators may not have the knowledge to provide high- quality annotations. |
| Approach: | They propose to directly model instance difficulty to improve model performance and to route instances to appropriate annotators. |
| Outcome: | The proposed model improves performance on a biomedical information extraction task using expert and lay annotations. |
Sentence-T5: Scalable Sentence Encoders from Pre-trained Text-to-Text Models (2022.findings-acl)
Copied to clipboard
| Challenge: | Sentence embeddings are useful for language processing tasks, but it is unclear how to produce them from encoder-decoder models. |
| Approach: | They investigate the effects of scaling up sentence encoders to 11B parameters on sentence embeddings from text-to-text transformers (T5) . |
| Outcome: | The proposed models outperform the previous best models on both SentEval and SentGLUE transfer tasks. |
Universal Sentence Encoder for English (D18-2)
Copied to clipboard
Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St. John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, Brian Strope, Ray Kurzweil
| Challenge: | TensorFlow Hub sentence embedding models have good task transfer performance . model variants allow for trade-offs between accuracy and compute resources . |
| Approach: | They propose easy-to-use TensorFlow Hub sentence embedding models with good task transfer performance. |
| Outcome: | The proposed models outperform models without transfer learning and those that use only word-level transfer on a number of NLP tasks. |
Universal Sentence Representation Learning with Conditional Masked Language Model (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to learn sentence representations on unlabeled corpora are difficult and expensive to obtain, making it hard to cover many domains and languages. |
| Approach: | They propose a method to train sentence representations on large unlabeled corpora by conditioning on the encoded vectors of adjacent sentences. |
| Outcome: | The proposed method outperforms existing models on SentEval and can be extended to a broad range of languages and domains. |
Crisscrossed Captions: Extended Intramodal and Intermodal Semantic Similarity Judgments for MS-COCO (2021.eacl-main)
Copied to clipboard
| Challenge: | Existing image captioning datasets have limited cross-modal associations, preventing researchers from examining how inter-modal learning impacts intra-modal tasks. |
| Approach: | They propose to use image captioning data to support multi-modal retrieval training and evaluation to assess the impact of inter-modality learning. |
| Outcome: | The proposed model is able to measure the influence of intra- and inter-modality learning. |
ReQA: An Evaluation for End-to-End Answer Retrieval Models (D19-58)
Copied to clipboard
| Challenge: | Popular QA benchmarks like SQuAD have driven progress on identifying answer spans within a specific passage . retrieving relevant answers from a huge corpus of documents is still a challenging problem . |
| Approach: | They propose a benchmark for evaluating large-scale sentence-level answer retrieval models . they establish baselines using both neural encoding models and classical retrieval techniques . |
| Outcome: | The proposed model outperforms human models on identifying answer spans within a specific passage . the proposed model is scalable and can bypass the typical document retrieval step . |
LongT5: Efficient Text-To-Text Transformer for Long Sequences (2022.findings-naacl)
Copied to clipboard
| Challenge: | Recent work has shown that increasing the input length or increasing model size can improve the performance of Transformer-based neural models. |
| Approach: | They propose a model that integrates attention ideas from long-input transformers and adopts pre-training strategies from summarization pre-train into the scalable T5 architecture. |
| Outcome: | The proposed model outperforms the original T5 models on several summarization and question answering tasks and achieves state-of-the-art results. |
LAReQA: Language-Agnostic Answer Retrieval from a Multilingual Pool (2020.emnlp-main)
Copied to clipboard
| Challenge: | LAReQA tests for “strong” cross-lingual alignment, requiring semantically related cross-language pairs to be closer in representation space than unrelated same-language pair. |
| Approach: | They propose a new benchmark for language-agnostic answer retrieval from a multilingual candidate pool that tests for "strong" cross-lingual alignment . they augment training data via machine translation and find that model performance is improved by augmenting training data through machine translation . |
| Outcome: | The proposed task is based on multilingual BERT (mBERT) and XLM-R. |
STAIR: Learning Sparse Text and Image Representation in Grounded Tokens (2023.emnlp-main)
Copied to clipboard
Chen Chen, Bowen Zhang, Liangliang Cao, Jiguang Shen, Tom Gunter, Albin Jose, Alexander Toshev, Yantao Zheng, Jonathon Shlens, Ruoming Pang, Yinfei Yang
| Challenge: | State-of-the-art contrastive learning models like CLIP and ALIGN are less interpretable and suffer from inferior accuracy than dense representations. |
| Approach: | They extend CLIP and ALIGN models to build a sparse semantic representation that is interpretable and easy to integrate with existing retrieval systems. |
| Outcome: | The proposed model outperforms CLIP and ALIGN models on image and text retrieval tasks with a 4.9% and +4.3% improvement on COCO-5k textimage and imagetext retrieval respectively. |
A Corpus with Multi-Level Annotations of Patients, Interventions and Outcomes to Support Language Processing for Medical Literature (P18-1)
Copied to clipboard
| Challenge: | In 2015 alone, about 100 manuscripts describing randomized controlled trials for medical interventions were published every day. |
| Approach: | They propose a corpus of 5,000 medical articles annotated with demarcations of text spans that describe the Patient population enrolled, the Interventions studied and to what they were Compared, and the Outcomes measured. |
| Outcome: | The proposed corpus includes 5,000 medical articles describing clinical randomized controlled trials. |
Large Dual Encoders Are Generalizable Retrievers (2022.emnlp-main)
Copied to clipboard
Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hernandez Abrego, Ji Ma, Vincent Zhao, Yi Luan, Keith Hall, Ming-Wei Chang, Yinfei Yang
| Challenge: | Experimental results show that dual encoders outperform sparse and dense retrievers on the BEIR dataset significantly. |
| Approach: | They challenge belief that bottleneck layer is too limited for out-of-domain generalization . they scale up the model while keeping bottleneck as a single dot-product with a fixed size . |
| Outcome: | The proposed model outperforms sparse and dense retrievers on the BEIR dataset significantly. |
SueNes: A Weakly Supervised Approach to Evaluating Single-Document Summarization via Negative Sampling (2022.naacl-main)
Copied to clipboard
| Challenge: | Existing studies on automatic summary evaluation metrics focus on lexical similarity and require a reference summary which is expensive to obtain. |
| Approach: | They propose to use a weakly supervised summary evaluation approach without the presence of reference summaries to transform existing summarization datasets into corrupted reference summarizers. |
| Outcome: | The proposed method outperforms baselines and shows that it improves linguistic quality over all metrics. |
Cross-Domain Review Helpfulness Prediction Based on Convolutional Neural Networks with Auxiliary Domain Discriminators (N18-2)
Copied to clipboard
| Challenge: | Recent studies on review helpfulness prediction require labeled samples for each domain/category of interest. |
| Approach: | They propose a convolutional neural network based model which leverages word-level and character-based representations to transfer knowledge between domains. |
| Outcome: | The proposed model outperforms the state-of-the-art on the Amazon product review dataset. |
CLIP-UP: A Simple and Efficient Mixture-of-Experts CLIP Training Recipe with Sparse Upcycling (2025.findings-emnlp)
Copied to clipboard
Xinze Wang, Chen Chen, Yinfei Yang, Hong-You Chen, Bowen Zhang, Aditya Pal, Xiangxin Zhu, Xianzhi Du
| Challenge: | Mixture-of-Experts (MoE) models are crucial for scaling model capacity while controlling inference costs. |
| Approach: | They propose an alternative training strategy that converts a dense CLIP model into a sparse MoE architecture. |
| Outcome: | The proposed training strategy outperforms dense models on COCO and Flickr30k benchmarks. |
On the Intractability to Synthesize Factual Inconsistencies in Summarization (2024.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for detecting factual inconsistencies in abstractive summarization are lacking in factual consistency detection. |
| Approach: | They propose to use real model-generated summaries with human annotations to detect factual inconsistencies. |
| Outcome: | The proposed model outperforms the SOTA on CoGenSumm, FactCC, Frank, and SummEval datasets. |
A Simple and Effective Method To Eliminate the Self Language Bias in Multilingual Representations (2021.emnlp-main)
Copied to clipboard
| Challenge: | Language agnostic and semantic-language information isolation is an emerging research direction for multilingual representations models. |
| Approach: | They propose a method that factors out language identity information from semantic related components in multilingual representations pre-trained on monolingual data. |
| Outcome: | The proposed method improves cross-lingual transfer performance on weak alignment models. |
MURAL: Multimodal, Multitask Representations Across Languages (2021.findings-emnlp)
Copied to clipboard
Aashi Jain, Mandy Guo, Krishna Srinivasan, Ting Chen, Sneha Kudugunta, Chao Jia, Yinfei Yang, Jason Baldridge
| Challenge: | Image-caption pairs and translation pairs provide the means to learn deep representations of and connections between languages. |
| Approach: | They propose a dual encoder that integrates image-text matching and translation pairs to solve two tasks by learning from billions of pairs. |
| Outcome: | The proposed encoder outperforms ALIGN's cross-modal retrieval performance on well-resourced languages and significantly improves on under-resource languages. |
DocAsRef: An Empirical Study on Repurposing Reference-based Summary Quality Metrics as Reference-free Metrics (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing reference-based metrics are limited by their reliance on human input. |
| Approach: | They propose to adapt some reference-based metrics to assess system summary against human-written references. |
| Outcome: | The proposed model outperforms reference-based metrics on two datasets and is comparable to reference-free metrics. |
Language-agnostic BERT Sentence Embedding (2022.acl-long)
Copied to clipboard
| Challenge: | Existing methods for learning bilingual sentence embeddings are not well explored. |
| Approach: | They propose to combine best methods for learning multilingual sentence embeddings with pre-trained models to achieve 83.7% bi-text retrieval accuracy over 112 languages on Tatoeba. |
| Outcome: | The proposed model achieves 83.7% bi-text retrieval accuracy over 112 languages on Tatoeba, above the 65.5% achieved by LASER. |
Syntactic Patterns Improve Information Extraction for Medical Search (N18-2)
Copied to clipboard
| Challenge: | Medical professionals search the literature by specifying the type of patients, the medical intervention(s) and the outcome measure(s). |
| Approach: | They propose to exploit the availability of structured abstracts to extract medically relevant information from syntactic patterns. |
| Outcome: | The proposed models differ from the constituent unigrams in the extracted patterns, suggesting that they capture contextual information that is otherwise lost. |
PAWS-X: A Cross-lingual Adversarial Dataset for Paraphrase Identification (D19-1)
Copied to clipboard
| Challenge: | Existing work on adversarial data generation focuses on English . Existing multilingual datasets show effectiveness of deep, multilingual pre-training . |
| Approach: | They propose a dataset of 23,659 human translated PAWS evaluation pairs in six languages . they show the effectiveness of deep, multilingual pre-training while leaving considerable headroom . |
| Outcome: | The proposed model shows that multilingual training and evaluation regimes are more accurate than previous models. |
Multi-stage Training with Improved Negative Contrast for Neural Passage Retrieval (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing neural firststage retrieval models overcome lexical gap issue by projecting query and document to a shared dense space. |
| Approach: | They propose a multi-stage framework for neural passage retrieval using synthetic data, negative sampling, and fusion techniques. |
| Outcome: | The proposed framework improves retrieval accuracy and enhances the negative contrast in both stages. |